基于知识蒸馏的分子性质预测及其可扩展性分析OACHSSCD

Knowledge distillation maintains predictive accuracy while reducing computational costs by transferring knowledge from complex teacher models to streamlined student models.This paper constructs a knowledge distil-lation framework based on a multi-task weighted optimization loss function specifically for molecular property pre-diction tasks.Using three graph neural network models,namely SchNet,DimeNet++,and TensorNet,the effectiveness of the framework in both domain-specific and cross-domain knowledge distillation settings was inves-tigated systematically.In the domain-specific scenario,teacher models trained on the QM9 dataset guide student models in predicting unseen quantum mechanical properties within the same dataset,verifying the potential of the proposed method for enhancing precision and efficiency.In the cross-domain scenarios,knowledge distillation is employed to transfer QM9-pretrained teacher embeddings to experimental datasets such as ESOL(solubility)and FreeSolv(hydration free energy),the predictive accuracy of the proposed method in cross-domain,small-sample training environments was demonstraed.The results indicate that knowledge distillation is a robust strategy for en-hancing molecular representation learning,providing critical technical support for the core requirements of high precision and low computational cost in materials science and drug discovery.